The research from Purdue University, first spotted by news outlet Futurism, was presented earlier this month at the Computer-Human Interaction Conference in Hawaii and looked at 517 programming questions on Stack Overflow that were then fed to ChatGPT.
“Our analysis shows that 52% of ChatGPT answers contain incorrect information and 77% are verbose,” the new study explained. “Nonetheless, our user study participants still preferred ChatGPT answers 35% of the time due to their comprehensiveness and well-articulated language style.”
Disturbingly, programmers in the study didn’t always catch the mistakes being produced by the AI chatbot.
“However, they also overlooked the misinformation in the ChatGPT answers 39% of the time,” according to the study. “This implies the need to counter misinformation in ChatGPT answers to programming questions and raise awareness of the risks associated with seemingly correct answers.”
GPT-2 came out a little more than 5 years ago, it answered 0% of questions accurately and couldn’t string a sentence together.
GPT-3 came out a little less than 4 years ago and was kind of a neat party trick, but I’m pretty sure answered ~0% of programming questions correctly.
GPT-4 came out a little less than 2 years ago and can answer 48% of programming questions accurately.
I’m not talking about mortality, or creativity, or good/bad for humanity, but if you don’t see a trajectory here, I don’t know what to tell you.
The study is using 3.5, not version 4.
4 produces inaccurate programming answers too
Obviously. But it is FAR better yet again.
Not really. I ask it questions all the time and it makes shit up.
Yes. But it is better than 3.5 without any doubt.
Lemmy seems to be very near-sighted when it comes to the exponential curve of AI progress, I think this is an effect because the community is very anti-corp
In what year do you estimating AI will have 90% accuracy?
No clue? Somewhere between a few years (assuming some unexpected breakthrough) or many decades? The consensus from experts (of which I am not) seems to be somewhere in the 2030s/40s for AGI. I’m guessing accuracy probably will be more on a topic by topic basis, LLMs might never even get there, or only related to things they’ve been heavily trained on. If predictive text doesn’t do it then I would be betting on whatever Yann LeCun is working on.